I build machine learning systems, and the teams that ship them. I'm the Engineering Lead at Srotas Health, where we build AI for healthcare: an agentic system that evaluates clinical criteria across 1B+ patient records in under ten seconds, LLM-backed search, and document pipelines for patient feasibility assessment.
Before that I led the ML engineering team at Fairmatic, built conversational AI at Sense, and went from backend to ML engineering at Helpshift. Eleven years in all, the last several on production ML, LLM pipelines and agentic systems.
🌐 pardeep-singh.com has the full story: experience, projects and what I've been up to lately.
I contribute mostly to AI infrastructure: observability, vector search, distributed compute and healthcare NLP.
| Project | What I work on | PRs |
|---|---|---|
| OpenMed · healthcare NLP and de-identification | ICD-10-CM, RxNorm and HPO linkers, FHIR builders, negation scope, synthetic HL7 v2 and C-CDA data with PHI ground truth, multilingual de-identification, privacy metrics; federated fine-tuning with differential privacy in review | 37 merged · 50 in review |
| Arize OpenInference · OpenTelemetry for AI | Span handling for Agno and OpenAI Agents, google-genai output limits, GenAI semantic conventions | 4 merged · 6 in review |
| Agent Orchestrator · fleets of coding agents | Worktree and session lifecycle fixes, tmux and Windows support, Claude Code and Codex adapters | 6 merged · 13 in review |
| Milvus · vector database | NGRAM regex planning, expression-rewriter optimisations, exact DISTINCT |
3 merged · 7 in review |
| Arize Phoenix · AI observability and evals | REST endpoint for token counts and costs, px CLI commands, Python client resources |
1 merged · 13 in review |
| Ray · distributed compute for AI | Ray Serve cancellation and shutdown, Ray Jobs submission, uv run parsing |
2 merged · 12 in review |
| scikit-learn | Examples for the PrecisionRecall and ConfusionMatrix displays | 2 merged |
Earlier (2015–2018): subspace support in clj_fdb, a Clojure port of Redis in Action, scripts for the Helpshift API, fixes to OAuth 2 in Action, and an AWS provider for Boxupp.
- Fine-tuning small LLMs for classification: Llama 3.2 1B, Gemma 2B and Gemma 2B on financial sentiment (Kaggle bronze medals), compared against Qwen3, Phi-2, Open LLaMA and RoBERTa Large.
- Nifty50 news RAG: retrieval-augmented Q&A over stock news with LangGraph, OpenAI and Chroma, built on my own news dataset.
- Cricket datasets: ball-by-ball data for the ICC Men's World Cup 2023, T20 World Cup 2024, IPL, WPL and more, downloaded over 15,000 times.
- study-group-hf-agents-course: the Hugging Face Agents course study group I ran with my IISc deep learning cohort.
- ml_classroom: notebooks from the machine learning sessions I ran at Helpshift.
- Earlier builds: a minimal SQLite clone in C and a parking lot service in Clojure on FoundationDB.
On Kaggle I'm an Expert in both Datasets and Notebooks.
- LLMs and agents: agentic systems, RAG, structured LLM output, text-to-SQL, LangGraph, agent frameworks, voice-agent NLU
- Fine-tuning and deep learning: PyTorch, PyTorch Lightning, Hugging Face Transformers, fine-tuning small LLMs (Llama, Gemma, Qwen, Phi), scikit-learn
- Healthcare AI: clinical NLP, de-identification, clinical coding (ICD-10-CM, RxNorm), FHIR
- Search and retrieval: Elasticsearch, ChromaDB, Milvus, hybrid retrieval
- MLOps and observability: MLflow, model deployment platforms, LLM tracing and evals with OpenTelemetry, Phoenix and OpenInference
- Data and infrastructure: Python, PySpark, Kafka, Postgres, MongoDB, AWS. In an earlier life, Clojure and Java.
pardeep-singh.com · LinkedIn · Kaggle · X · Strava